Algorithmic Agency
Introduction
What is agency? In this essay I will explore how agency is conceptualized in the social sciences and what key debates and philosophical backgrounds have informed the current understanding of agency. Agency will be disambiguated from free will and choice in order to explore complex social phenomena such as algorithms and whether algorithmic agency is a useful and legitimate concept for alignment researchers to consider. In doing so, we will dissect the assumptions underlying the prevailing philosophical understanding of agency and point to its common pitfalls. These shortcomings will then be addressed with my own theories of agency which synthesize ideas from Quantum Physics (Barad), Economics (Ostrom), Cybernetics (Hayles), and Sociology (Bourdieu and Latour).
Short history of the agency versus structure debate
In the social sciences there is a longstanding debate over the priority of structure or agency in shaping human behaviour. Structure is most commonly defined as the recurrent patterned arrangements which influence or limit the choices (ability and options) available. Those factors of influence typically include social divisions such as gender, class, race, religion, ethnicity, nationality, ability, customs and more, that determine or limit decisions. Agency is traditionally defined as the ability of individual humans to act independently from these structures and to make their own “free” choices. This convention of defining agency contingently against structure, will be dissected to reveal some unquestioned assumptions about how agency has been defined in humanist and naturalist worldviews. Sociological understandings of agency will be cross-pollinated with ideas from economics and cybernetics in order to reach a fuller understanding of agency that can better account for algorithmic agency.
Short history of the free will versus determinism debate
Before engaging with contemporary ideas of agency, it is important to understand the more ancient theological debate between free will and determinism which forms the backdrop to the relatively more recent debates about agency and structure. Free will has been the central moral problem from thinkers as diverse as the Greek philosophical tradition, with Plato and Aristotle's debates about free will, to the Christian dilemma over the primacy of predestination versus free will in the debates between Calvinist and Arminian theologies, to the Islamic debates about free will between the Mutazilites versus the Ash'arites. At stake is nothing less than the moral responsibility to assign blame and credit. The concept of free will as paradigmatic of agency can be traced back to the European Enlightenment. Emerging from this period was the notion of the Liberal Humanist Subject (Hayles 1999) whose free will was his most defining feature. According to Liberal Humanism, the rational and free will of man is the most defining feature of agency. This long-standing assumption will be challenged by thinkers such as Latour, as we shall see. Thinking within this Humanist tradition, Karl Marx argued that in modern society, people were controlled by the ideologies of the bourgeoisie, Friedrich Nietzsche argued that man made choices based on his own selfish desires, or the "will to power" and, Freud accounted for the unconscious as an internal force that determines human behavior. As we shall see, these notions of control, choice and determination orbit the central notion of free will as paradigmatic of agency and why a fuller account of agency will necessitate breaking free from this orbit. Only by generating enough escape velocity can we break free from the gravitational influence of Liberal Humanism to embark on a new trajectory toward algorithmic agency.
What is free will?
Although there is a rich source of literature that seeks to characterize, define and systematize the many notions of free will that exist not only in philosopher’s minds but also in lay understanding, these debates shall not be addressed within the scope of this essay for the purposes of keeping on track with agency. To summarize these debates, free will can be defined in one of five ways:
Free will1: Ability to make choices (Dennett 1984b)
Free will2: Ability to make choices according to wants (Nahmias 2006)
Free will3: Ability to make choices according to wants independent from social structure (Clark 1998)
Free will4: Ability to make choices according to wants independent from external and internal constraints (Feldman 2017)
Free will5: Ability to make choices according to wants independent from causation (Lavazza 2016)
Free will shall be stipulatively defined for the purposes of this essay to mean free will3, free will4 and free will5 and shall be indicated as such by the superscript. External constraints entailed by free will4 include factors such as environment, society, nature, God/s, and other human agents; while internal constraints can include such factors as uncontrollable urges, needs, genes, and personality (Feldman 2017). The notion of free will as it has developed according to the Liberal Humanist tradition follows the trajectory that is captured by the third and fourth senses of free will3 and free will4 and less frequently by free will5. However, Liberal Humanist free will in this essay shall refer to all three of these definitions unless stated otherwise.
Choice also requires disambiguation as it can refer to one of three senses:
Choice1: a power, privilege, right, faculty, capacity, or ability to choose
Choice2: a process of choosing; choice-making
Choice3: a thing that is chosen; a choosable; an option
Choice throughout this essay shall be stipulatively defined as the process of choosing and shall refer exclusively to choice2 in order to avoid conflation with choice1 (ability to choose, also known as free will1) and choice3 (thing that is chosen).
Finally, choose shall be stipulatively defined as a process of consciously causing one thing in preference to another. Choosing shall be contrasted with decision-making which is stipulatively defined as a process of causing one thing in preference to another, regardless of whether it was done consciously or not. This distinction between conscious choosing and nonconscious decision-making is crucial for our understanding of algorithmic agency.
What is agency?
In order to answer the question of what agency is, we must first understand what agency is for. Agency is ascribed to certain causes because we deem these causes and their effects to be ethically relevant. Agency is considered essential to legal, moral, and political judgments (Juth & Lorentzon, 2010; Searle, 2007). To illustrate this concept, let us imagine a vignette: a falling durian fruit can land on the head of a person, thus killing them without the durian tree being morally culpable for their death. However, another person who releases such a durian from such a height and causes the death of someone would be morally responsible and be subject to legal charges of either manslaughter or even murder, depending on whether proof of intentionality can be established. The end result is the same for both cases: the death of a person. However, the ethical implications are vastly different. Therefore, agency is a property we recognize in certain effects whose causes have moral, legal and social relevance. The ethically relevant question is whether we can attribute blame or credit to the cause. In other words, is the cause in question blameworthy or praiseworthy?
In current sociological discourse, there are three key figures whose intellectual contributions toward our understanding of agency should be acknowledged. Here, we shall explore how agency is expounded by Anthony Giddens, George Herbert Mead, and Margaret Archer. According to Giddens, agency is the capacity of an individual to observe their experience and give reasons for their action. Giddens, therefore, identifies agency with reasoning and knowledge (Turker 1998). According to Mead, abstract philosophical and psychological concepts such as agency should be defined in terms of actual behaviour. A strict focus on behaviour helps in avoiding metaphysical impasses when dealing with such complex issues as decision-making and choice (Baldwin 1988). Therefore, Mead identifies agency with the behaviour of the individual. Margaret Archer on the other hand, builds on pre-existing ideas of structure and agency by developing them through her social realist approach. Archer does this by refining the notions of structure and agency into an analytical dualism which demands the study of the interplay between the two levels without conflating them in what she calls upward conflation of Gidden’s approach and downward conflation of Mead’s approach. Therefore, in Archer’s approach, her concept of the internal conversation in which agents' ability to enter into reflective conversation with themselves is foregrounded as a key feature of agency.
These three accounts of agency are all wrong because they uncritically identify agency with one or more aspects of the Liberal Humanist notion of free will, be it Giddens’ “reasoning and knowledge” or Mead's “behaviour of the individual” or even Archer's “analytical dualism” and “internal conversation”. Much of sociology inherits this Liberal Humanist bias of privileging the conscious intentionality of the thinking human subject’s agency over other forms of agency. Two thinkers who buck this trend, however, are Bruno Latour with his Actor-Network Theory and Pierre Bourdieu with his concept of the habitus, as we shall see later. As such, neither Giddens’ nor Mead’s nor Archer’s theories can account for nonconscious agency also known as nonintentional agency, a vital ingredient sorely missing if we are to include the relevance of algorithms in our enquiry. Another glaringly obvious shortcoming of Liberal Humanism is its inability to account for nonhuman agency. By focusing only on human consciousness, Liberal Humanist subjectivity ignores any consideration for embodied cognition and the vast social, historical, cultural, biological, technological, environmental, and political context which gives rise to human consciousness and in which consciousness is always embedded. Liberal Humanist free will, therefore, puts too much emphasis on human consciousness, thereby, narrowing the scope of agency only to a subject precariously perched on the razor thin edge of individual consciousness that is defined in antagonism to the wider context in which it is situated. The prototypical bearer of moral responsibility according to the philosophical tradition of Liberal Humanism is human (as opposed to a nonhuman animal, machine, god, alien, or corporation), free (as opposed to slave), and an adult (as opposed to juvenile). This narrow construction of agency is known in analytical jurisprudence as Juridical Humanism (another subset of Liberal Humanism) which is the subject of Western legal systems and was shaped by specific historical developments particular to Europe and does not hold true for other societies (Pietrzykowski 2018). Juridical Humanism and Liberal Humanism place undue importance on the primacy of free will as the defining notion of agency. Experimental research led by Renatas Berniūnas (2020) has demonstrated that the concept of free will has no cross-culturally universal conceptual content; that is to say, the concept of Liberal Humanist free will (in the sense of free will3, free will4 and free will5) is not universal but is instead specific to Western, Educated, Industrialized, Rich, and Democratic (WEIRD) societies and has no purchase in the minds of Chinese, Indian, Mongolian and Malay societies (Berniūnas 2020). This notion of agency, codified in the legal system and cognitively entrenched in the minds of WEIRD, societies is, therefore, not intuitive in other human societies, and inadequate for our current social needs. I argue that agency includes a much wider range of phenomena than free will and this expanded view encompasses algorithmic agency.
Nonintentional agency
One of the clearest examples that illustrate how agency is a much wider concept than free will is in the Common Law legal concepts of negligence and transferred intent which predate Liberal Humanism. Negligence is a concept in tort law stipulating that the harm caused by failing to act — whether intentional or unintentional — as a form of carelessness is still the responsibility of the agent regardless of their intent to freely will or choose the harm. Transferred intent is another legal doctrine that holds that, when the intention to harm one individual inadvertently causes another person to be hurt instead, the perpetrator is still held responsible. In both cases, the free will1 of the perpetrator was irrelevant since their choice2 — conscious, intentional, purposeful, and deliberate decision-making — was not implicated in the process. Yet, despite the lack of free will1, Common Law stipulates that the perpetrator has agency in this situation and, thus, is morally responsible for the harm caused by the action. These examples of nonintentional agency being recognized and enshrined in the Common Law legal system demonstrates how intentionality (an inalienable component of free will1) can be decoupled from agency. Therefore, nonintentional actions can be accommodated under the wider concept of agency in a way that free will cannot.
In contrast to the notion of nonintentional agency, the philosophy of Liberal Humanism (and its jurisprudential branch, Juridical Humanism) permits no recognition of an agency or free will separate from intentionality. For thinkers working within the Western philosophical tradition such as the Liberal Humanists but especially classical sociology, agency is believed to be legitimately exercised only when independent from external constraints and internal constraints. Hence, the structure versus agency dichotomy. This notion of agency as equivalent to free will4 is peculiar to Liberal Humanism since agency in classical sociology is not only distinguished from external causes such as structure, culture, tradition, society and other external causal factors but it is also distinguished from internal causes such as genetics, hormones, neurotype, and unconscious mental conditions that influence a moral agent's actions and choices. Freedom of will — according to the logical extreme of Liberal Humanism — must be free from any form of determinism to be considered truly free. This of course sends agency into retreat, occupying less and less of the causal fabric in order to be free from the very thing that makes agency possible in the first place. To quote Kant:
“The light dove, cleaving the air in her free flight, and feeling its resistance, might imagine that its flight would be still easier in empty space”.
If agency is uninfluenceable by cause-and-effect, it cannot cause effects in the world. Once agency is defined to be free from the sequence of causation, it renders agency impotent to enact changes in the world. By taking the logic of Liberal Humanist free will to its extreme (i.e. free will5), we arrive at the absurd notion that no one has agency. But just because no one has free will according to Liberal Humanism, does not mean that we can do away with moral responsibility and the notion of agency. What needs to be done is to sever the concept of agency from its Liberal Humanist orbit in order to account for the moral responsibility of agents who are not human and not conscious: algorithmic agency.
Nonhuman agency
Recognizing the anthropocentric bias of Liberal Humanism, the French sociologist and philosopher Bruno Latour was one of the first thinkers to locate agency neither in human “subjects” nor in nonhuman “objects”, but in “heterogeneous associations of humans and nonhumans”. Latour introduced Actor-Network Theory where agency is conceptualized as “the ability to enact difference”. Latour demonstrates this with the example of a gun-human collective. In contrast to Liberal Humanist proponents of the second amendment of the US constitution, who claim that “guns don't kill people, people kill people”, Latour argues that neither guns nor people are the agents responsible in whole for gun violence but rather the gun-person collective they form (Latour 1999, 193). In doing so, Latour grants agency to nonhuman entities and is the first to recognize forms of agency that are independent not only from intentionality and consciousness but also from humans. In doing so, Latour introduced two developments to the way social scientists theorize agency. The first is that agency need not presuppose intentionality. This distinction between agency and intentionality is important because it recognizes the notions of willed causes and unwilled causes as being the moral responsibility of the agent as exemplified in the legal doctrines of negligence and transferred intent. Agency in Actor-Network Theory does not presuppose intentionality nor does it presuppose consciousness. The second development of ANT is the distinction between human agency and nonhuman agency. The recognition that agency is not exclusive to humans but is distributed amongst humans and nonhumans was an idea that was popularized by Latour and later developed by John Law. Taken together, the recognition of nonhuman agency and nonintentional agency allows us to encompass a wider range of phenomena under the scope of sociological and moral consideration beyond the limiting paradigm of Liberal Humanist free will (Hayles 2017). Building upon Latour’s ANT, Karen Barad expands the concept of agency to explore the ethical and ontological implications of taking ANT agency to its extreme conclusion in their notion of Agential Realism which conceptualizes agency in its widest sense as “being in its intraactivity” (Barad 2007). Barad extends the scope of their Agential Realist account of agency to include not just the social reality of humans but goes further to include the nonhuman material reality of subatomic particles as having agency (Barad 2007, p. 66). Everything has agency because everything has the ability to enact material change in the world. Barad also argues that every instance of material change has some degree of ethical or normative relevance, however small it may be. In doing so, Barad questions the long-held dualisms of subject-object, human-nonhuman, mental-physical, social-material that exist to sustain the Liberal Humanist worldview. These discussions, as interesting as they are, are beyond the scope of this essay and so will not be discussed for the sake of brevity and keeping on track with agency.
Thus far, we have two ways of conceptualizing agency. On one end of the spectrum, agency is defined as free will5 which limits agency to an ever vanishingly thin sliver of disembodied rational consciousness — the subject of Liberal Humanism, supposedly free from external and internal constraints and ultimately, from all causation. On the other hand, we have the ANT notion of agency as “the ability to enact difference” which, in its mature Agential Realist development, is so all-encompassing as to include everything from guns to subatomic particles as having agency. With these two extreme ways of conceptualizing agency, we can now focus on the relevant level in order to fashion a serviceable theory to understand algorithmic agency.
Nonconscious cognitive agency
Bruno Latour ushered in the posthuman turn when he proposed Actor-Network Theory in the early 1980s in which he locates agency “neither in humans nor in nonhumans but in heterogeneous associations of humans and nonhumans”. However, in attributing agency to nonhuman causes, Latour also makes the mistake of anthropomorphizing nonhumans. For instance, Latour depicts how the “mode of technological beings […] misunderstands” other modes of existence (2013: p215). Furthermore, Latour does very little to recognize the different degrees of agency that reside in different agents. In my view, Latour is correct to recognize the agency of nonhumans but in assuming the flat ontology of all nonhumans and conceptualizing their agency in anthropocentric ways, he fails to distinguish the different degrees of agency between mountains, plants, nonhuman animals and humans. What is needed then, is a more sophisticated “categorization of agencies” that has greater analytical precision (Kipnis 2015; p49).
Laidlaw attempts to address these concerns as below:
“When considering agency, cause, intention, state, and response are matters of concern. To be blamed for something, an individual or group of individuals must be held to have caused a negative situation; their blame increases if this result was intended by the actor; their blame might be reduced if their state of mind was induced by an outside actor—a spirit who possessed them, a temporary madness that overtook them or perhaps drugs or alcohol; finally, the actor should have the capacity to make a response. An agent without the ability to pay compensation, or who lives outside the jurisdiction where the relevant authorities can capture him, is simply not worth pursuing.” (Laidlaw 2002; p315)
Like Laidlaw (2002), I agree with the notion that blame is not an all-or-nothing state but rather can be conceptualized in degrees. Different entities can have different scales of agency to reflect their differing capabilities to enact causes in the world. In such a model, and in contrast to Latour's flat ontology, a conscious human has more agency than a dog, which in turn has more agency than a plant, which in turn has more agency than a rock. Indeed, Kipnis argues for the relevance of consciousness in ethical concerns.
Kipnis himself suggests:
“Granting agency to everything (which is to acknowledge that anything and everything could affect us) should include differentiating types of agency. The types of agency we get with life and self-consciousness are particularly worth differentiating and highlighting. Self-consciousness raises the problems of ethics and choice, and our own ethical concerns require us to consider the effects of our agency on other humans above all.” (Kipnis 2015; 55)
In bringing human consciousness back into the picture, we must not privilege the anthropocentric notion of free will over other forms of agency, especially nonhuman entities such as algorithms. In this respect, Latour and the other sociologists of the ANT school do not dedicate enough consideration to dissecting the notions of thought, consciousness, cognition, intentionality and decision-making sufficiently enough to adequately account for algorithmic agency.
Algorithms are part of what N Katherine Hayles calls the cognitive nonconscious. Drawing upon discoveries in the cognitive sciences, Hayles explains how the existence of nonconscious cognition are processes that are inaccessible to conscious introspection but nevertheless are essential for consciousness to function. Such examples of nonconscious but nonetheless cognitive processes include the low-level cognition of brainstem function which maintains basic life processes such as homeostasis, heart rate regulation, temperature, wake and sleep cycles, and digestion which occur below the awareness of a conscious self. Nonconscious cognition also includes cybernetic technologies capable of decision-making without the input of conscious choice-making. In her book Unthought, Hayles (2017) proposes a tripartite framework of cognition in which analytical distinctions are drawn between material processes, nonconscious cognitive material processes, and conscious cognitive material processes. Consciousness, Hayles argues, occupies a central position in our thinking due to the importance that Liberal Humanism places on free will as a rational and conscious decision-making process. This overemphasis on conscious cognition overshadows the role of cognition that is nonconscious. Cognition, by contrast, is a much broader capacity that extends far beyond consciousness into other neurological brain processes. Thus, cognition is also pervasive in other life forms and complex technical systems. In contrast to the notion of free will, cognition is a process embodied in rather than disembodied from materiality. However, just because cognition is embedded in material processes does not mean it is equivalent to it. The crucial distinguishing characteristics of cognition that separates it from the underlying material processes are (conscious) choice-making and (nonconscious) decision-making, and thus possibilities for interpretation and meaning. Cognition can thus be defined as a process that interprets information within contexts that connect it with meaning (Hayles 2017). Cognition encompasses consciousness but is itself differentiated from non-cognitive material processes, contra ANT’s flat ontology. Traditionally, the humanities have been concerned with meanings relevant to humans in human-dominated contexts. The framework developed by Hayles challenges that orientation, insisting cognitive processes happen within a broad spectrum of possibilities that include nonhuman animals and plants as well as technical systems. By widening the notion of cognition beyond merely consciousness, nonconscious cognitive processes can include medical diagnostic systems, automated satellite imagery identification, ship navigation systems, weather prediction programs, and a host of other nonconscious cognitive devices that interpret ambiguous or conflicting information to arrive at conclusions that are rarely completely certain. Furthermore, by limiting cognition to interpretation of information, Hayles makes a valuable distinction between cognizers and material processes such as avalanches, tsunamis, tornados, blizzards, sandstorms, and hurricanes which have incredible agency but lack the capacity for decision-making and therefore can only perform as (non-cognitive) agents, not as (cognizing) actors embedded in cognitive assemblages with moral and ethical implications. Actor is, therefore, a term that is reserved for a cognizer while agent corresponds to material forces and objects to signify their lesser degree of agency. This latter category includes objects that may act as cognitive supports (similar to the analogous supporting role that the brainstem plays for consciousness); it also includes material forces that may be harnessed to perform cognitive tasks when suitable constraints are introduced, for example, when electrical voltages are transformed into a bit stream within a computational medium or when a calculator is used to aid in an accountant’s calculations.
With regard to algorithms, Hayles qualifies a useful caveat:
"a computer algorithm, written as instructions on paper, is not itself cognitive, for it becomes a process only when instantiated in a platform capable of understanding the instruction set and carrying it out. That interprets information: interpretation implies a decision. There must be more than one option for interpretation to operate" (Hayles 2017; 40)
Latour and Hayles make the case for recognizing the nonhuman agency of algorithms, however, Hayles' caveat reminds us that algorithms are not only embedded in material processes but that those material processes are embedded in wider institutional arrangements which can only be accounted for by shifting gears to the macrosocial level. Not only do we need a cognitive level, we also need an institutional level of analysis to account for algorithmic agency.
Institutional agency: Pierre Bourdieu versus Elinor Ostrom
In order to understand theories of institutions and how they shape behaviour, we must first understand the notion of habitus introduced by Pierre Bourdieu. The work of Pierre Bourdieu is an attempt to reconcile structure and agency, as external structures are internalized into the habitus while the actions of the agent externalize interactions between actors into the social relationships in the field. Bourdieu's theory, therefore, is a dialectic between “externalizing the internal”, and “internalizing the external”. Clever as this development may be, in order for this theory to work, Bourdieu introduces terms such as capital, field, and practice which are vague and imprecise to bolster his definition of habitus. Such terms lend themselves more to hermeneutic speculation of their “true” meaning than to empirical testing. For instance, the difference between field and practice is fuzzy and makes these Bourdieusian concepts less amenable to empirical investigation in a way that could have been avoided if he had operationalized their definitions more rigorously in a sharply delineated formal system. Due to this lack of rigour in his loosely defined concepts, Bourdieusian concepts suffer from what Harvey Whitehouse calls the hermeneutic vortex (2003):
In the social sciences it is customary to spill much ink trying to wrest control of the meaning of its terms of art from other theoreticians, as if defining a word were equivalent to providing an explanation or making an intellectual contribution. Many proceed as if the meaning of a word was something to be discovered through careful application of quasi-philosophical reflection, and the philosopher was the professional whose task was to discover it.
To rescue Bourdieu's habitus as a useful notion in which the external is internalized (i.e. the value-prioritization system of a structure can be internalized into the value-prioritization system of an agent), it is better to draw from another thinker, the economist Elinor Ostrom whose ground-breaking work in developing the theoretical foundations of a Grammar of Institutions contributed to her winning the 2009 Nobel prize in economics.
Ostrom’s syntax of institutions is able to account for Bourdieu's concept of habitus with greater clarity by considering habitus not as a "structured and structuring structure" but rather as a set of institutional statements which can be classified into 3 types: rules, norms, and conventions. These in turn can be analysed according to her ADICO syntax. ADICO is an acronym for the components that make up all institutional statements (otherwise known as habitus in Bourdieusian terms). All institutional statements can be broken down into 5 "grammatical" components: ATTRIBUTES, DEONTIC, AIM, CONDITIONS, and OR-ELSE. Attribute is the actor to which the statement applies; Deontic tells if an action is permitted, required, or forbidden; Aim is generally a verb that describes the goal of the institutional statement and the action to which the deontic refers; Condition is when or where an action may, must, or must not take place; and Or-else is the consequence or sanction of not following a rule. Depending on the number of components present, the ADICO syntax is able to account for all types of institutional statements such as conventions, norms and rules.
All conventions can be written as [ATTRIBUTES][AIM][CONDITIONS] (AIC); all norms can be written as [ATTRIBUTES][DEONTIC][AIM][CONDITIONS] (ADIC); all rules can be written as [ATTRIBUTES][DEONTIC][AIM][CONDITIONS][OR-ELSE] (ADICO). The syntax is cumulative: norms contain all of the components of a convention plus a DEONTIC; rules contain all the components of a norm plus an OR-ELSE.
Institutional statements refer to a shared linguistic constraint or opportunity that prescribes, permits or advises actions or outcomes for actors (both individual and corporate). Institutional statements are coded, spoken, written, or tacitly understood in a form intelligible to actors in an empirical setting. In theoretical analyses, institutional statements will often be interpretations or abstractions of empirical constraints and opportunities — what has traditionally been conceived of as structure. The syntax of the grammar operationalizes the structural descriptions it identifies. The advantage that institutional grammar offers over Bourdieu's habitus is that the ADICO syntax can be applied not only at an institutional level but also at an individual level, therefore, the value-prioritization system of the institution becomes aligned with the value-prioritization system of the individual. In Bourdieusian terms, this would be known as “internalizing the external”. However, because the ADICO syntax offers clarity with regard to the specific values of the deontic, aim, attributes, conditions and or-else consequences, an individual has the power to decide whether the value-prioritization system of the institution aligns with their own and consciously make choices on whether to conform to it or to change one of the components of the syntax in accordance with their own value-prioritization system. In Bourdieusian terms, this would be known as “externalizing the internal”.
We now have all the ingredients to consider the agency of algorithms
What are algorithms?
Algorithms are such a ubiquitous part of our daily decision-making. Algorithms are responsible for the information we see on our Facebook feeds, for the kinds of video recommendations we get on YouTube, and even financial markets are dominated by High Frequency Trading (HFT) algorithms. But what is an algorithm? Simply put, an algorithm is a sequence of well-defined instructions to solve a problem, perform calculations, process data, conduct automated reasoning and — most relevantly for our discussion — make decisions for us. Algorithms are most commonly rendered in a coding language that can be read by computers.
The Facebook algorithm
In order to conduct a sociological analysis of the Facebook algorithm, we must first understand its historical development. The Facebook Algorithm is not a single algorithm as is commonly believed. It has undergone multiple reincarnations to arrive at its most current form. The first Facebook algorithm was created in 2009 with the purpose of sorting the order for newsfeeds based on each post's popularity; this was in stark contrast to the much simpler reverse-chronological order of news presentation. This version of the algorithm valued popularity as defined according to certain metrics over other considerations when choosing what to put on display. The next modifications to the algorithm came during a decade of further tinkering: personalized feeds that analyse tens of thousands of data points in order to maximize the amount of time spent on the platform. This tweaking of the Facebook algorithm came about as a result of prioritizing the interests of shareholders who incentivized more screen time because of higher exposure to ads. The reasoning behind this was that greater exposure to ads would result in more consumer spending of advertised products and therefore functioned primarily to maximize profit. In 2016, in a shift of tone, Facebook began prioritizing posts from friends and family, as well as “informative” and “entertaining” content. The algorithms started to take into account the amount of time users spent with a post, even if they didn't like or share the post. Live videos were also prioritized. However, in January 2018, responding in part to widespread criticism, Mark Zuckerberg announced Facebook newsfeed changes that prioritize posts that spark conversations and meaningful interactions. The change was meant to increase the quality, rather than the quantity, of the time that people spend on Facebook, as well as take more responsibility for how the platform affects its users’ well-being. As of 2020, representatives of Facebook have stated that their focus is on helping users understand the algorithm, and take control of those ranking signals to give it better feedback. The algorithm currently decides for the user by ranking the post they see in the order that they're likely to enjoy them, based on a variety of factors known as ranking signals. Ranking signals are data points about a user's past behaviour and the behaviour of everyone else on the platform too.
What can be demonstrated through this fine-grained discussion of the Facebook Algorithm is the very social concerns that influence the development of the algorithm. Different priorities were foregrounded at different stages of the Facebook Algorithm's evolution. Rather than conceptualizing the algorithm as an “objective” and impartial calculation of data, the Facebook Algorithm prioritizes different ranking signals to align with the values of the relevant social actors. It is clear from the history of the Facebook algorithm that social actors play a vital role in influencing the decision-making prioritization system of the algorithm. At different points in the lifetime of the algorithm, different human values such as popularity, time, shareholder interests, friends and family, or meaningful interactions were prioritized over others. These choices were determined by very social concerns whose values fluctuate according to sociological factors. Human choices are institutionalized, entrenching these values which then become coded into the decision-making process of the algorithms which then have a serious impact on the cognition of Facebook’s users. Edwin Hutchins developed the concept of “distributed cognition” because, according to this view: “human cognition is not just influenced by cultures and society, it is in a very fundamental sense a cultural and social process” (Hutchins 1995). Cognition is a process that interprets information within contexts that connect it with meaning. As we have seen, the context is an impenetrable morass of algorithmic and institutional decisions amidst a quagmire of human choices.
Conclusion
To summarize, we have seen that agency is a concept for assigning blame and credit to an agent. Agency in the Liberal Humanist paradigm is equated with free will because of the primacy it places on the conscious and intentional choice of human agents. This notion of free will fails to account for the agency of nonintentional actors in various legal situations as well as the agency of nonhumans in political and ethical disputes such as in the case of the gun-human collective. Through Actor-Network Theory, we expand our notion of agency to include all material processes that have a causal relationship to human action. By fine tuning agency to focus on cognizing actors as having a greater degree of agency than mere material processes, we are able to single out the agency of cognition without necessarily foregrounding human consciousness at the expense of the nonconscious cognition of algorithms. In recognizing that algorithms are processes embodied in wider technological and institutional contexts, we carve out a space for analysing the distributed agency of algorithms as an instantiation of institutional statements. With a better understanding of the conscious, social and human choices that underlie the decision-making process of institutions and ultimately of algorithms, perhaps we can better account for the algorithmic agency of the algorithm-institution-human collective to make decisions that serve the benefit of everyone.
Reference List
Baldwin, J. D. (1988) ‘Habit, Emotion, and Self-Conscious Action’, Sociological perspectives: SP: official publication of the Pacific Sociological Association, 31(1), pp. 35–57.
Barad, K. (2007) Meeting the Universe Halfway: Quantum Physics and the Entanglement of Matter and Meaning. Duke University Press.
Berniūnas, R. et al. (2020) ‘The weirdness of belief in free will’, Consciousness and cognition, 87, p. 103054.
Bourdieu, P. (1990) ‘Structures, habitus, practices’, The logic of practice, pp. 52–65.
Clark, W. R. (1998) ‘Agents and Structures: Two Views of Preferences, Two Views of Institutions’, International studies quarterly: a publication of the International Studies Association, 42(2), pp. 245–270.
Crawford, S. E. S. and Ostrom, E. (1995) ‘A Grammar of Institutions’, The American political science review, 89(3), pp. 582–600.
Dennett, D. C. (1984) ‘I Could not have Done Otherwise--So What?’, The journal of philosophy, 81(10), pp. 553–565.
Feldman, G. (2017) ‘Making sense of agency: Belief in free will as a unique and important construct: Belief in free will construct review’, Social and personality psychology compass, 11(1), p. e12293.
How the Facebook Algorithm Works in 2020 and How to Work With It (2020). Available at: https://blog.hootsuite.com/facebook-algorithm/ (Accessed: 11 January 2021).
Juth, N. and Lorentzon, F. (2010) ‘The concept of free will and forensic psychiatry’, International journal of law and psychiatry, 33(1), pp. 1–6.
Katherine Hayles, N. (2017) Unthought: The Power of the Cognitive Nonconscious. University of Chicago Press.
Kipnis, A. (2015) ‘Agency between humanism and posthumanism: Latour and his opponents’, HAU: Journal of Ethnographic Theory, 5(2), pp. 43–58.
Laidlaw, J. (2002) ‘For an Anthropology of Ethics and Freedom’, The journal of the Royal Anthropological Institute, 8(2), pp. 311–332.
Latour, B. (2011) ‘Network Theory| Networks, Societies, Spheres: Reflections of an Actor-network Theorist’, International Journal of Communication Systems, 5(0), p. 15.
Latour, B. and Centre de Sociologie de L’Innovation Bruno Latour (1999) Pandora’s Hope: Essays on the Reality of Science Studies. Harvard University Press.
Lavazza, A. (2016) ‘Free Will and Neuroscience: From Explaining Freedom Away to New Ways of Operationalizing and Measuring It’, Frontiers in human neuroscience, 10, p. 262.
Monroe, A. E. and Malle, B. F. (2010) ‘From Uncaused Will to Conscious Choice: The Need to Study, Not Speculate About People’s Folk Concept of Free Will’, Review of philosophy and psychology, 1(2), pp. 211–224.
Nahmias, E. et al. (2005) ‘Surveying Freedom: Folk Intuitions about free will and moral responsibility’, Philosophical psychology, 18(5), pp. 561–584.
Pietrzykowski, T. (2018) ‘Personhood Beyond Humanism’, SpringerBriefs in Law. doi: 10.1007/978-3-319-78881-4.
Searle, J. R. (2007) ‘Dualism revisited’, Journal of physiology, Paris, 101(4-6), pp. 169–178.
Tucker, K. (1998) Anthony Giddens and Modern Social Theory. SAGE.